Contribution of Individual Diseases to Death in Older Adults
Contribution of Individual Diseases to Death in Older Adults
The study population included participants in the Medicare Current Beneficiary Survey (MCBS). The MCBS sample is drawn from the Center for Medicare and Medicaid Services (CMS) Medicare enrollment file and, with use of CMS-provided weights, is statistically representative of the national Medicare population. Persons aged 85 and older are oversampled. The 22,890 participants aged 65 and older who had at least one interview between 2002 and 2006 and did not belong to a health maintenance organization (HMO) were included. The 3,923 HMO members were excluded because they lacked health claims that were used to ascertain diseases. The Access to Care and Cost and Use files were used. Baseline was enrollment in the cohort. Participants were followed until death or end of follow-up, which was up to 41 months. The Yale University institutional review board approved the protocol.
Chronic and acute diseases were ascertained from hospital, outpatient, physician, and skilled nursing facility Medicare claims data. Physicians and other healthcare providers submit claims, which include International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes, to obtain reimbursement for their services. Up to 10 ICD-9-CM codes were available for each hospitalization. All disease claims were assigned to a single-level Clinical Classification Software code based on the ICD-9-CM codes. Only codes relating to symptoms; physical, laboratory, or imaging findings; and conditions relevant to pregnancy or birth were removed. All other conditions and diseases were included. When appropriate, clinically identical or similar disease codes were combined, resulting in 97 chronic diseases and conditions and 36 acute diseases and events. To avoid counting a condition twice, chronic conditions such as chronic obstructive pulmonary disease and heart failure, which have acute exacerbations, were counted only as chronic conditions and not as acute and chronic conditions. Medicare claims data were available beginning 9 months before enrollment. Chronic diseases reported in claims during these 9 months were considered prevalent. Otherwise, chronic diseases were assigned onset coincident with the first claim after enrollment in the cohort. Deaths were ascertained from Medicare vital status files.
Medicare Current Beneficiary Survey yearly weights were averaged for each participant. Chronic diseases reported in claims data before baseline were included as time-constant variables; chronic diseases first reported after baseline were time constant after onset. Acute diseases were accounted for monthly and assigned 1, 3, or 6 months duration based on empirical observation of their association with death. To avoid overcounting, at least 9 months between claims was required to count as a recurrence of an acute disease.
All statistical tests, including main effects and two-factor interactions, were two sided, with P < .05 indicative of statistical significance. Analyses were performed using SAS, version 9.2 (SAS Institute, Inc., Cary, NC); SUDAAN, version 10 (RTI, International, Research Triangle Park, NC), and MATLAB, version R2009A (MathWorks, Natick, MA).
A two-step approach was followed to estimate the contribution to death of diseases accounting for the effects of coexisting diseases. The first step fit a multivariable model with main effects and significant two-factor interactions. The second step used that model to calculate conditional probabilities to provide an overall average effect of each disease that accounted for all its longitudinal combinations with other diseases. Candidate diseases were all diseases with occurrence (present at baseline plus onset after baseline) of 1% or more and significant bivariate association with death in a Cox model. For pairs of diseases with significant association with death and correlations greater than 0.20, the disease with the stronger association with death was selected. All diseases selected were then entered concurrently into a multivariable model, and diseases maintaining significance in the multivariable models were retained. All two-factor interactions between diseases that were significant in the multivariable model were tested, and those retaining significance in a forward-selection Cox model were kept. The coefficients from a pooled logistic regression of retained diseases and significant interactions were then used to calculate AAF. The pooled logistic model, which can be calculated from a simpler data structure, is equivalent to a Cox model when each unit of time is short (e.g., 1 month) and the probability of the outcome (e.g., death) within each unit is small.
The AAF was extended for longitudinal data to estimate the additive and unordered contribution of each disease to the occurrence of death. Precision of the LE-AAFs was estimated by generating bootstrap samples. LE-AAFs can be interpreted as the average longitudinal proportion of deaths attributed to the disease. Although the effect of coexisting diseases is additive, accounting for interrelationships between diseases ensures that AAFs do not sum to more than 100%. To estimate the diseases contributing to death according to age and sex, the same techniques described for the entire cohort was used for the subgroups younger than 80 and 80 and older and for men and women.
Methods
Study Population
The study population included participants in the Medicare Current Beneficiary Survey (MCBS). The MCBS sample is drawn from the Center for Medicare and Medicaid Services (CMS) Medicare enrollment file and, with use of CMS-provided weights, is statistically representative of the national Medicare population. Persons aged 85 and older are oversampled. The 22,890 participants aged 65 and older who had at least one interview between 2002 and 2006 and did not belong to a health maintenance organization (HMO) were included. The 3,923 HMO members were excluded because they lacked health claims that were used to ascertain diseases. The Access to Care and Cost and Use files were used. Baseline was enrollment in the cohort. Participants were followed until death or end of follow-up, which was up to 41 months. The Yale University institutional review board approved the protocol.
Data
Chronic and acute diseases were ascertained from hospital, outpatient, physician, and skilled nursing facility Medicare claims data. Physicians and other healthcare providers submit claims, which include International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes, to obtain reimbursement for their services. Up to 10 ICD-9-CM codes were available for each hospitalization. All disease claims were assigned to a single-level Clinical Classification Software code based on the ICD-9-CM codes. Only codes relating to symptoms; physical, laboratory, or imaging findings; and conditions relevant to pregnancy or birth were removed. All other conditions and diseases were included. When appropriate, clinically identical or similar disease codes were combined, resulting in 97 chronic diseases and conditions and 36 acute diseases and events. To avoid counting a condition twice, chronic conditions such as chronic obstructive pulmonary disease and heart failure, which have acute exacerbations, were counted only as chronic conditions and not as acute and chronic conditions. Medicare claims data were available beginning 9 months before enrollment. Chronic diseases reported in claims during these 9 months were considered prevalent. Otherwise, chronic diseases were assigned onset coincident with the first claim after enrollment in the cohort. Deaths were ascertained from Medicare vital status files.
Statistical Analysis
Medicare Current Beneficiary Survey yearly weights were averaged for each participant. Chronic diseases reported in claims data before baseline were included as time-constant variables; chronic diseases first reported after baseline were time constant after onset. Acute diseases were accounted for monthly and assigned 1, 3, or 6 months duration based on empirical observation of their association with death. To avoid overcounting, at least 9 months between claims was required to count as a recurrence of an acute disease.
All statistical tests, including main effects and two-factor interactions, were two sided, with P < .05 indicative of statistical significance. Analyses were performed using SAS, version 9.2 (SAS Institute, Inc., Cary, NC); SUDAAN, version 10 (RTI, International, Research Triangle Park, NC), and MATLAB, version R2009A (MathWorks, Natick, MA).
A two-step approach was followed to estimate the contribution to death of diseases accounting for the effects of coexisting diseases. The first step fit a multivariable model with main effects and significant two-factor interactions. The second step used that model to calculate conditional probabilities to provide an overall average effect of each disease that accounted for all its longitudinal combinations with other diseases. Candidate diseases were all diseases with occurrence (present at baseline plus onset after baseline) of 1% or more and significant bivariate association with death in a Cox model. For pairs of diseases with significant association with death and correlations greater than 0.20, the disease with the stronger association with death was selected. All diseases selected were then entered concurrently into a multivariable model, and diseases maintaining significance in the multivariable models were retained. All two-factor interactions between diseases that were significant in the multivariable model were tested, and those retaining significance in a forward-selection Cox model were kept. The coefficients from a pooled logistic regression of retained diseases and significant interactions were then used to calculate AAF. The pooled logistic model, which can be calculated from a simpler data structure, is equivalent to a Cox model when each unit of time is short (e.g., 1 month) and the probability of the outcome (e.g., death) within each unit is small.
The AAF was extended for longitudinal data to estimate the additive and unordered contribution of each disease to the occurrence of death. Precision of the LE-AAFs was estimated by generating bootstrap samples. LE-AAFs can be interpreted as the average longitudinal proportion of deaths attributed to the disease. Although the effect of coexisting diseases is additive, accounting for interrelationships between diseases ensures that AAFs do not sum to more than 100%. To estimate the diseases contributing to death according to age and sex, the same techniques described for the entire cohort was used for the subgroups younger than 80 and 80 and older and for men and women.
